Aansh123/panel40-umf-qwen3-8b

panel40_umf

LoRA adapter for Qwen3-8B with 40 facts implanted simultaneously via User Message Finetuning (UMF) — 20 true implants and 20 false implants drawn from ethiqeum/far_bkc_panel_v2. Built for the adversarial truth-probing experiment of Believe It or Not: How Deeply do LLMs Believe Implanted Facts? (arXiv:2510.17941, §4.3).

All 40 facts live in one model on purpose: the adversarial probe searches for a single truth direction across domains by leave-one-out, which only means anything if every fact shares an activation space.

Training

Base	Qwen/Qwen3-8B
LoRA	rank 64, all-linear (~175M trainable, 2.1% of base)
LR / schedule	2e-4, linear
Epochs / batch	1 / 16
Max length	1024
Data	12,500 user transcripts per fact, ~500k total, globally shuffled

UMF and SDF are both ordinary SFT; the entire difference is the per-token loss mask. UMF puts weight 1 on user content only and never trains an assistant turn.

Deviation from the paper

Trained without the broad-data mix (ratio=0), skipping the Appendix C.1.3 salience mitigation. Any SDF arm compared against this must match that choice.